noether.core.utils.training¶

Submodules¶

Classes¶

`UpdateCounter`	UpdateCounter is used to track the current update, epoch, and sample
`ScheduleWrapper`	Wrapper around a schedule that handles getting the value based on an UpdateCounter and whether the schedule is
`TrainingIteration`

Package Contents¶

class noether.core.utils.training.UpdateCounter(start_iteration, end_iteration, updates_per_epoch, effective_batch_size)¶

UpdateCounter is used to track the current update, epoch, and sample number during training.

Initializes UpdateCounter.

Parameters:

start_iteration (noether.core.utils.training.training_iteration.TrainingIteration) – The iteration (measured in either epochs, updates, or samples) at which the training will start. If the training is resumed the values will be updated accordingly.
end_iteration (noether.core.utils.training.training_iteration.TrainingIteration) – The iteration (measured in either epochs, updates, or samples) at which the training will end.
updates_per_epoch (int) – Number of updates per epoch.
effective_batch_size (int) – Effective batch size.

updates_per_epoch¶

start_iteration¶

end_iteration¶

cur_iteration¶

effective_batch_size¶

property is_full_epoch: bool¶

Returns True if the current update is a full epoch.

Return type:: bool

property epoch_as_float: float¶

Returns the current epoch as a float value.

Return type:: float

property epoch: int | None¶

Returns the current epoch.

Return type:: int | None

property update: int | None¶

Returns the current update.

Return type:: int | None

property sample: int | None¶

Returns the current sample.

Return type:: int | None

property is_finished: bool¶

Returns True if the end checkpoint is reached.

Return type:: bool

next_epoch()¶

Increments the current epoch by 1.

Return type:: None

next_update()¶

Increments the current update by 1.

Return type:: None

add_samples(num_samples)¶

Adds samples to the current sample count.

Parameters:: num_samples (int)
Return type:: None

class noether.core.utils.training.ScheduleWrapper(schedule, update_counter=None, interval='update')¶

Wrapper around a schedule that handles getting the value based on an UpdateCounter and whether the schedule is based on updates or epochs.

Parameters:

schedule (noether.core.schedules.ScheduleBase) – The ScheduleBase to wrap.
update_counter (noether.core.utils.training.counter.UpdateCounter | None) – The UpdateCounter to use for getting the current step. If None, the schedule is assumed to be constant.
interval (Literal['update', 'epoch']) – Whether the schedule is based on updates or epochs. Interval should be either “update” or “epoch”.

schedule¶

update_counter = None¶

interval = 'update'¶

get_value()¶

Get the current value of the schedule based on the current step in the UpdateCounter.

Return type:: float

class noether.core.utils.training.TrainingIteration(epoch=None, update=None, sample=None)¶

Parameters:

epoch (int | None)
update (int | None)
sample (int | None)

epoch = None¶

update = None¶

sample = None¶

copy()¶

Creates a copy of a TrainingIteration object.

Return type:: TrainingIteration

property specified_properties_count: int¶

Counts the number of specified (non-None) properties.

Return type:: int

property is_fully_specified: bool¶

Checks if all properties are specified.

Return type:: bool

property is_minimally_specified: bool¶

Checks if at least one property is specified.

Return type:: bool

get_n_equal_properties(other)¶

Counts the number of equal properties between two TrainingIteration objects.

Parameters:: other (TrainingIteration) – The other TrainingIteration object to compare with.
Return type:: int

to_fully_specified(updates_per_epoch, effective_batch_size)¶

Converts the TrainingIteration object to a fully specified TrainingIteration object. This method calculates the missing properties based on the specified ones.

Parameters:

updates_per_epoch (int) – The number of updates per epoch.
effective_batch_size (int) – The effective batch size used in the training process.

Return type:

TrainingIteration

classmethod from_dict(data)¶

Parameters:: data (dict[str, int])
Return type:: TrainingIteration

has_same_specified_properties(other)¶

Checks if the specified properties of two TrainingIteration objects are the same.

Parameters:: other (TrainingIteration) – The other TrainingIteration object to compare with.
Return type:: bool

static from_string(value)¶

Converts a training iteration string into a TrainingIteration object.

The training iteration string should be in the format “E{epoch}_U{update}_S{sample}”. For instance: E5_U12_S123 corresponds to TrainingIteration(epoch=5, update=12, sample=123).

Parameters:: value (str) – The TrainingIteration string to convert.
Return type:: TrainingIteration

static contains_string(source)¶

Checks if the source string contains a training iteration string.

Parameters:: source (str) – The source string to check.
Returns:: True if the source string contains a training iteration string, False otherwise.
Return type:: bool

static find_string(source)¶

Checks if the source string contains a training iteration string and returns it.

Parameters:: source (str) – The source string to check.
Returns:: The found TrainingIteration string.
Return type:: str

static from_filename(fname)¶

Creates a TrainingIteration object from a filename that contains a training iteration string.

Parameters:: fname (str) – The filename to extract the training iteration string from.
Returns:: The created object.
Return type:: TrainingIteration

static to_fully_specified_from_filenames(directory, training_iteration, prefix=None, suffix=None)¶

Converts a minimally specified TrainingIteration to a fully specified one from filenames in a given folder. The first file containing a training iteration string and optionally starting with prefix or ending with suffix is used.

Parameters:

directory (str) – The directory containing training iteration files.
training_iteration (TrainingIteration) – The TrainingIteration object to match against.
prefix (str | None) – Optional prefix to filter files.
suffix (str | None) – Optional suffix to filter files.

Returns:

The created object.

Return type:

TrainingIteration

Raises:

FileNotFoundError – If no matching file is found.

to_target_specification(target)¶

Creates a new object that matches the target specification.

Example: ` self=TrainingIteration(epoch=6, update=12, sample=123)` target=TrainingIteration(epoch=5) print(self.to_target_specification(target)) TrainingIteration(epoch=6) ` :param target: The target TrainingIteration object to match against.

Returns:: The created object.
Return type:: TrainingIteration
Parameters:: target (TrainingIteration)